consequently incorrect patient treatment [241]. Training the classifier with large and diverse data and avoiding overfitting and underfitting the model can help develop a generalizable classifier [280]. The developed classifier should also have balanced results and avoid high false-positive and false-negative rates to avoid overdiagnosis or underdiagnosis of patients [107,281]. Open access repositories are a good source of large and diverse data that can be used to increase the accuracy and generalizability of the classifier, and benchmark it [282]. These repositories could save the hassle of acquiring medical images and all the legal and technical issues involved. In this thesis, the open access datasets of the PROMISE12 [202] and PROSTATEx [203] challenges were used in Paper I and Paper II. It is a good practice to separate the training, validation, and testing sets of the classifier [283]. This separation can help avoid the risk of overfitting, increase the estimation accuracy of the model performance and improve the generalizability of the classifier [283]. One possible approach to train more generalizable classifiers with variant data from different countries and institutions is federated learning.
Federated learning performs at the client level, so there is no need to transfer the data to have it in one place [284,285]. In federated learning, each client (e.g., institution/hospital) can train the classifier on its own data. The weights of the classifier are then uploaded to a server that is shared by all clients, so that the other clients can download the weights and continue the training process of the classifier with their data [284,285].
5.8 Research ethics, data management and privacy aspects
The use of medical images to develop AI-based systems, e.g., CAD, is subject to a specific procedure including collecting ethical approvals, data access, querying data, data de-identification, data transfer and storage, QC, structuring data, and labelling data [286]. Medical images and all supporting clinical data are considered sensitive data. Data that should not be collected without ethical approvals [286]. Approvals are usually granted by institutional and/or local ethics committees before the study begins. Ethics committee members evaluate the benefits, harms, and risks of the systems to be developed [286]. For most medical image analysis studies, informed or passive consent must be obtained from patients [286]. Researchers and developers of AI systems have an obligation to respect the dignity and rights of patients [287]. Patient data should be secured, not sold, and not used in a way that contradicts the ethical consent given [287]. In this thesis, the ethical perspective was explored before starting the work.
The ethical approvals to collect and use the in-house collected dataset were obtained from the Regional Committee for Medical and Health Research Ethics (REC Mid Norway; identifiers 2013/1869 and 2017/576), and signed informed consent was obtained from patients. Patients
59
retained the right to withdraw consent to the use of their images at any time during the study period. To ensure the patients involvement, the thesis work was discussed during the users’
panel meetings with the researchers at the MR Cancer group at the Norwegian University of Science and Technology. The panel (https://www.ntnu.edu/isb/mr-cancer-user-involvement), which was established in close dialogue with the Norwegian Cancer Society, includes four former patients, two breast cancer and two prostate cancer patients, who provide important insights and participate in ethical and scientific discussions.
Medical images are typically accessed, queried and retrieved through Image Archiving and Communication System (PACS) [287]. Each medical institution has its own PACS, which requires access permission, granted after reviewing the access request and ensuring that ethical approval has been obtained. Locating the desired data in PACS is a tedious and time-consuming process that must be done carefully. Researchers should not view data for which they have not granted access permission. In this thesis, these guidelines were carefully followed and the permission to access and retrieve data from PACS was granted by the department of Radiology and Nuclear Medicine, St. Olavs Hospital, Trondheim University Hospital, Trondheim, Norway.
Before using the data to develop the CAD system, the data should be de-identified [286]. This includes anonymizing/pseudonymizing the metadata of the images and renaming the cases [288,289]. If a list is needed to link patients to their new pseudonymized identifiers, this list should be carefully stored in a different location than the data. Once de-identified, the data can be transferred to a secure storage point for later use in developing the system [286]. In this thesis, the European General Data Protection Regulation (GDPR) act [290] was followed, the data was pseudonymized and uploaded to a secure server on HUNT Cloud [291] and the link list was stored securely in a different location. HUNT Cloud is am ISO-certified digital infrastructure that allows data controllers and researchers to store, access and analyse sensitive data in controlled environments. HUNT Cloud is in compliance with the European GDPR and Norwegian acts and regulations for research and data security.
After transferring the data to a secure server, it should be systematically structured and checked for quality before being used for AI system development to avoid inherited errors [286]. All or part of the data should be labelled by experts to provide references for AI system development and evaluation [286]. All these aspects have been considered and followed in this work.
60
To ensure a good data management protocol in this thesis, the data and the AutoRef and segmentation QC code was treated according to the FAIR (Findability, Accessibility, Interoperability, and Reusability) principles [292]. AutoRef and the segmentation QC code have been made publicly available on GitHub (https://github.com/ntnu-mr-cancer) with clear instructions for their use
61
6 Conclusions and future perspectives
This thesis aimed to facilitate the integration of automated CAD systems of prostate cancer using mpMRI into clinical practice by developing and evaluating new image pre-processing, segmentation and quality control methods to improve the performance of the CAD workflow.
CAD systems have the potential to overcome many of the pitfalls of traditional prostate cancer diagnostics. Especially when integrated with mpMRI, which provides multiple anatomical and functional parameters and quantitative information that can improve the diagnostic process.
CAD usually consists of a chain of steps, which implement ML-based methods to achieve a specific task. Each step depends on the performance of the previous step, i.e., if one of the steps fails or commits an error, the following steps are prone to propagate that error, potentially leading to misdiagnosis. Therefore, the implemented methods should be automated, accurate and transparent. The work in this thesis focused on the early steps of the CAD workflow, in particular the T2W MRI normalization and prostate segmentation, as ensuring high performance and error control of these steps reduces the risk of propagated errors. This could not only improve the performance of CAD, but also increase the confidence of the radiologists in these systems.
T2W MR images require normalization of signal intensity to allow quantitative analysis, which is the direction CAD and related statistical models follow. Several normalization methods have been proposed for this purpose, but with limitations. In this thesis a new dual-reference tissue normalization approach that automatically extracts the signal intensity of the fat and muscle around the prostate to normalize the image was proposed. To the best of our knowledge, this is the first multi-reference tissue normalization approach where the delineation of ROIs is fully automated. The proposed method was found to increase the intensity homogeneity between patients and within patients scanned multiple times. Moreover, it showed better performance than other commonly used normalization methods. The method was also shown to improve classification between healthy and malignant prostate tissue in PZ and non-PZ. The proposed method is generalizable, transparent, easy to implement and made publicly available.
Another important step in the CAD workflow is VOI segmentation, in this case of the prostate.
This step, which defines the VOIs to be used later for feature extraction, could be efficiently performed automatically using DL-based methods. Despite the overall good performance, these methods can still produce poor segmentation masks in some cases, which calls for a QC step. In this thesis, a generalizable, transparent, publicly available segmentation QC system
62
was developed. The system assigns a score to each segmentation related to its quality, which can be used to distinguish between acceptable and poor segmentations. This system is an important step towards implementing DL-based segmentation methods in clinical practice and reducing human intervention.
DL-based segmentations could also be used in clinical applications that rely on multiple scans in time, such as for patients on active surveillance. Therefore, it is extremely important that the segmentations generated by DL-based methods are not only accurate but also reproducible. In this thesis, the reproducibility of DL-based segmentations of the prostate and its zones was investigated. The reproducibility of the best-performing DL-based methods were found to be comparable to that of manual segmentations.
In conclusion, this thesis shows that the performance of the early steps of automated CAD for prostate cancer can be improved and controlled, leading to more generalizable, transparent and trustworthy systems. This is seen as an important step towards the integration of CAD systems into clinical practice.
This thesis could be fundamental for further research to improve robust, generalizable and transparent CAD systems for prostate cancer. Normalization can be further improved by developing new methods that build on AutoRef and perhaps incorporate additional reference tissues. The segmentation QC system can be extended to include prostate zones segmentation.
Developing a similar QC system for mpMRI registration would be helpful and may improve the performance of CAD. Conducting studies to investigate the reproducibility of the various radiomics features and pre-processing steps would be very informative and would provide useful suggestions to extract features correctly. CAD systems have the potential to decide whether or not biopsy sampling is necessary, help detect the suspicious areas and help targeting them when biopsy sampling is necessary. Despite the existence of several CAD systems aimed at detecting or grading prostate cancer, there is still room for the development of more robust and trustworthy systems. In the era of open science, these systems should benefit from previous research and methods developed for the various CAD steps to ensure better performance than the previous systems. However, the most important step for the future is to test the various CAD systems in the clinic and ensure that they meet the radiologists' expectations. This is crucial for building a trust relationship between the radiologists and the CAD systems, which will hopefully lead to the actual implementation of CAD in the clinic.
63
7 Bibliography
1. Stratton, M.R.; Campbell, P.J.; Futreal, P.A. The cancer genome. Nature 2009, 458, 719-724, doi:10.1038/nature07943.
2. Hanahan, D.; Weinberg, R.A. The hallmarks of cancer. Cell 2000, 100, 57-70, doi:10.1016/s0092-8674(00)81683-9.
3. Hanahan, D.; Weinberg, R.A. Hallmarks of cancer: the next generation. Cell
2011, 144, 646-674, doi:10.1016/j.cell.2011.02.013.4. Bray, F.; Ferlay, J.; Soerjomataram, I.; Siegel, R.L.; Torre, L.A.; Jemal, A.
Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2018, 68, 394-424, doi:10.3322/caac.21492.
5. Siegel, R.L.; Miller, K.D.; Jemal, A. Cancer statistics, 2020. CA Cancer J Clin
2020, 70, 7-30, doi:10.3322/caac.21590.6. Shah, R.B.; Zhou, M. Prostate Biopsy Interpretation: An Illustrated Guide;
Springer Berlin Heidelberg: 2012; pp. 1 online resource.
7. Lee, C.H.; Akin-Olugbade, O.; Kirschenbaum, A. Overview of prostate anatomy, histology, and pathology. Endocrinol Metab Clin North Am 2011, 40, 565-575, viii-ix, doi:10.1016/j.ecl.2011.05.012.
8. Kierszenbaum, A.L.; Tres, L.L. Histology and cell biology : an introduction to pathology, 3rd ed.; Elsevier Saunders: Philadelphia, 2012; pp. xiv, 701 p.
9. Verma, S.; Rajesh, A. A clinically relevant approach to imaging prostate cancer:
review. AJR Am J Roentgenol 2011, 196, S1-10 Quiz S11-14, doi:10.2214/AJR.09.7196.
10. Canadian Cancer Society. The prostate. Availabe online:
https://www.cancer.ca/en/cancer-information/cancer-type/prostate/prostate-cancer/the-prostate (accessed on 05.Jan.2021).
11. Hall, J.E.; Guyton, A.C. Guyton and Hall textbook of medical physiology, Thirteenth edition. ed.; Elsevier: Philadelphia, PA, 2016; pp. xix, 1145 pages.
12. Kumar, V.; Abbas, A.K.; Aster, J.C.; Robbins, S.L. Robbins basic pathology, Tenth edition. ed.; Elsevier: Philadelphia, Pennsylvania, 2018; pp. xiv, 935 pages.
13. Cancer Registry of Norway. Cancer in Norway 2019 - Cancer incidence, mortality, survival and prevalence in Norway. Oslo: Cancer Registry of Norway
2020.14. Pernar, C.H.; Ebot, E.M.; Wilson, K.M.; Mucci, L.A. The Epidemiology of Prostate Cancer. Cold Spring Harb Perspect Med 2018, 8, doi:10.1101/cshperspect.a030361.
15. Perdana, N.R.; Mochtar, C.A.; Umbas, R.; Hamid, A.R. The Risk Factors of Prostate Cancer and Its Prevention: A Literature Review. Acta Med Indones
2016, 48, 228-238.16. Gann, P.H. Risk factors for prostate cancer. Rev Urol 2002, 4 Suppl 5, S3-S10.
17. Steinberg, G.D.; Carter, B.S.; Beaty, T.H.; Childs, B.; Walsh, P.C. Family
history and the risk of prostate cancer. Prostate
1990,17, 337-347,
doi:10.1002/pros.2990170409.
64
18. Peisch, S.F.; Van Blarigan, E.L.; Chan, J.M.; Stampfer, M.J.; Kenfield, S.A.
Prostate cancer progression and mortality: a review of diet and lifestyle factors.
World J Urol 2017, 35, 867-874, doi:10.1007/s00345-016-1914-3.
19. Calle, E.E.; Rodriguez, C.; Walker-Thurmond, K.; Thun, M.J. Overweight, obesity, and mortality from cancer in a prospectively studied cohort of U.S.
adults. N Engl J Med 2003, 348, 1625-1638, doi:10.1056/NEJMoa021423.
20. Miller, D.C.; Hafez, K.S.; Stewart, A.; Montie, J.E.; Wei, J.T. Prostate carcinoma presentation, diagnosis, and staging: an update form the National Cancer Data Base. Cancer 2003, 98, 1169-1178, doi:10.1002/cncr.11635.
21. Leslie, S.W.; Soon-Sutton, T.L.; Sajjad, H.; Siref, L.E. Prostate Cancer. In StatPearls, Treasure Island (FL), 2020.
22. Saad, F.; Clarke, N.; Colombel, M. Natural history and treatment of bone complications in prostate cancer. Eur Urol 2006, 49, 429-440, doi:10.1016/j.eururo.2005.12.045.
23. Barentsz, J.O.; Richenberg, J.; Clements, R.; Choyke, P.; Verma, S.; Villeirs, G.;
Rouviere, O.; Logager, V.; Futterer, J.J.; European Society of Urogenital, R.
ESUR prostate MR guidelines 2012. Eur Radiol 2012, 22, 746-757, doi:10.1007/s00330-011-2377-y.
24. Mottet, N.; Bellmunt, J.; Bolla, M.; Briers, E.; Cumberbatch, M.G.; De Santis, M.; Fossati, N.; Gross, T.; Henry, A.M.; Joniau, S., et al. EAU-ESTRO-SIOG Guidelines on Prostate Cancer. Part 1: Screening, Diagnosis, and Local Treatment with Curative Intent. Eur Urol 2017, 71, 618-629, doi:10.1016/j.eururo.2016.08.003.
25. Nadler, R.B.; Humphrey, P.A.; Smith, D.S.; Catalona, W.J.; Ratliff, T.L. Effect of inflammation and benign prostatic hyperplasia on elevated serum prostate specific antigen levels. J Urol 1995, 154, 407-413, doi:10.1097/00005392-199508000-00023.
26. Thompson, I.M.; Pauler, D.K.; Goodman, P.J.; Tangen, C.M.; Lucia, M.S.;
Parnes, H.L.; Minasian, L.M.; Ford, L.G.; Lippman, S.M.; Crawford, E.D., et al.
Prevalence of prostate cancer among men with a prostate-specific antigen level
< or =4.0 ng per milliliter. N Engl J Med 2004, 350, 2239-2246, doi:10.1056/NEJMoa031918.
27. McGuire, B.B.; Helfand, B.T.; Loeb, S.; Hu, Q.; O'Brien, D.; Cooper, P.; Yang, X.; Catalona, W.J. Outcomes in patients with Gleason score 8-10 prostate cancer:
relation to preoperative PSA level. BJU Int 2012, 109, 1764-1769, doi:10.1111/j.1464-410X.2011.10628.x.
28. Loeb, S.; Catalona, W.J. What is the role of digital rectal examination in men undergoing serial screening of serum PSA levels? Nat Clin Pract Urol 2009, 6, 68-69, doi:10.1038/ncpuro1294.
29. Kirby, R. The role of PSA in detection and management of prostate cancer.
Practitioner 2016, 260, 17-21, 13.
30. Bangma, C.H.; Roemeling, S.; Schroder, F.H. Overdiagnosis and overtreatment of early detected prostate cancer. World J Urol 2007, 25, 3-9, doi:10.1007/s00345-007-0145-z.
31. Helsedirektør. Nasjonalt handlingsprogram med retningslinjer for diagnostikk,
behandling og oppfølging av prostatakreft: NASJONALE FAGLIGE
65
RETNINGSLINJER. Helsedirektoratet, Avdeling spesialisthelsetjenester, Oslo
2020.32. U. S. Preventive Services Task Force; Grossman, D.C.; Curry, S.J.; Owens, D.K.; Bibbins-Domingo, K.; Caughey, A.B.; Davidson, K.W.; Doubeni, C.A.;
Ebell, M.; Epling, J.W., Jr., et al. Screening for Prostate Cancer: US Preventive Services Task Force Recommendation Statement. JAMA 2018, 319, 1901-1913, doi:10.1001/jama.2018.3710.
33. Naji, L.; Randhawa, H.; Sohani, Z.; Dennis, B.; Lautenbach, D.; Kavanagh, O.;
Bawor, M.; Banfield, L.; Profetto, J. Digital Rectal Examination for Prostate Cancer Screening in Primary Care: A Systematic Review and Meta-Analysis.
Ann Fam Med 2018, 16, 149-154, doi:10.1370/afm.2205.
34. Wallner, L.; Frencher, S.; Hsu, J.W.; Loo, R.; Huang, J.; Nichol, M.; Jacobsen, S. Prostate cancer screening trends in a large, integrated health care system. Perm J 2012, 16, 4-9.
35. Gosselaar, C.; Kranse, R.; Roobol, M.J.; Roemeling, S.; Schroder, F.H. The interobserver variability of digital rectal examination in a large randomized trial for the screening of prostate cancer. Prostate
2008,68, 985-993, doi:10.1002/pros.20759.
36. Cui, T.; Kovell, R.C.; Terlecki, R.P. Is it time to abandon the digital rectal examination? Lessons from the PLCO Cancer Screening Trial and peer-reviewed literature. Curr Med Res Opin 2016, 32, 1663-1669, doi:10.1080/03007995.2016.1198312.
37. Ilic, D.; Neuberger, M.M.; Djulbegovic, M.; Dahm, P. Screening for prostate cancer. Cochrane Database Syst Rev 2013, 10.1002/14651858.CD004720.pub3, CD004720, doi:10.1002/14651858.CD004720.pub3.
38. Donovan, J.; Hamdy, F.; Neal, D.; Peters, T.; Oliver, S.; Brindle, L.; Jewell, D.;
Powell, P.; Gillatt, D.; Dedman, D., et al. Prostate Testing for Cancer and Treatment (ProtecT) feasibility study. Health Technol Assess 2003, 7, 1-88, doi:10.3310/hta7140.
39. Hambrock, T.; Hoeks, C.; Hulsbergen-van de Kaa, C.; Scheenen, T.; Futterer, J.;
Bouwense, S.; van Oort, I.; Schroder, F.; Huisman, H.; Barentsz, J. Prospective assessment of prostate cancer aggressiveness using 3-T diffusion-weighted magnetic resonance imaging-guided biopsies versus a systematic 10-core transrectal ultrasound prostate biopsy cohort. Eur Urol 2012, 61, 177-184, doi:10.1016/j.eururo.2011.08.042.
40. Kvale, R.; Moller, B.; Wahlqvist, R.; Fossa, S.D.; Berner, A.; Busch, C.;
Kyrdalen, A.E.; Svindland, A.; Viset, T.; Halvorsen, O.J. Concordance between Gleason scores of needle biopsies and radical prostatectomy specimens: a population-based study. BJU Int 2009, 103, 1647-1654, doi:10.1111/j.1464-410X.2008.08255.x.
41. Jayadevan, R.; Zhou, S.; Priester, A.M.; Delfin, M.; Marks, L.S. Use of MRI-ultrasound Fusion to Achieve Targeted Prostate Biopsy. J Vis Exp 2019, 10.3791/59231, doi:10.3791/59231.
42. Moore, C.M.; Robertson, N.L.; Arsanious, N.; Middleton, T.; Villers, A.; Klotz,
L.; Taneja, S.S.; Emberton, M. Image-guided prostate biopsy using magnetic
66
resonance imaging-derived targets: a systematic review. Eur Urol 2013, 63, 125-140, doi:10.1016/j.eururo.2012.06.004.
43. Murphy, I.G.; NiMhurchu, E.; Gibney, R.G.; McMahon, C.J. MRI-directed cognitive fusion-guided biopsy of the anterior prostate tumors. Diagn Interv Radiol 2017, 23, 87-93, doi:10.5152/dir.2016.15445.
44. Weinreb, J.C.; Barentsz, J.O.; Choyke, P.L.; Cornud, F.; Haider, M.A.; Macura, K.J.; Margolis, D.; Schnall, M.D.; Shtern, F.; Tempany, C.M., et al. PI-RADS Prostate Imaging - Reporting and Data System: 2015, Version 2. Eur Urol 2016, 69, 16-40, doi:10.1016/j.eururo.2015.08.052.
45. Mottet, N.; van den Bergh, R.C.N.; Briers, E.; Van den Broeck, T.; Cumberbatch, M.G.; De Santis, M.; Fanti, S.; Fossati, N.; Gandaglia, G.; Gillessen, S., et al.
EAU-EANM-ESTRO-ESUR-SIOG Guidelines on Prostate Cancer-2020 Update. Part 1: Screening, Diagnosis, and Local Treatment with Curative Intent.
Eur Urol 2021, 79, 243-262, doi:10.1016/j.eururo.2020.09.042.
46. Gleason, D.F.; Mellinger, G.T. Prediction of prognosis for prostatic adenocarcinoma by combined histological grading and clinical staging. J Urol
1974, 111, 58-64, doi:10.1016/s0022-5347(17)59889-4.47. Pierorazio, P.M.; Walsh, P.C.; Partin, A.W.; Epstein, J.I. Prognostic Gleason grade grouping: data based on the modified Gleason scoring system. BJU Int
2013, 111, 753-760, doi:10.1111/j.1464-410X.2012.11611.x.48. Chen, N.; Zhou, Q. The evolving Gleason grading system. Chin J Cancer Res
2016, 28, 58-64, doi:10.3978/j.issn.1000-9604.2016.02.04.49. Epstein, J.I.; Egevad, L.; Amin, M.B.; Delahunt, B.; Srigley, J.R.; Humphrey, P.A.; Grading, C. The 2014 International Society of Urological Pathology (ISUP) Consensus Conference on Gleason Grading of Prostatic Carcinoma:
Definition of Grading Patterns and Proposal for a New Grading System. Am J Surg Pathol 2016, 40, 244-252, doi:10.1097/PAS.0000000000000530.
50. Buyyounouski, M.K.; Choyke, P.L.; McKenney, J.K.; Sartor, O.; Sandler, H.M.;
Amin, M.B.; Kattan, M.W.; Lin, D.W. Prostate cancer - major changes in the American Joint Committee on Cancer eighth edition cancer staging manual. CA Cancer J Clin 2017, 67, 245-253, doi:10.3322/caac.21391.
51. Nketiah, G.A. Magnetic Resonance Imaging for Improved Prostate Cancer Diagnosis. Doctoral thesis. Norwegian University of Science and Technology (NTNU), Trondheim, 2018.
52. Filson, C.P.; Marks, L.S.; Litwin, M.S. Expectant management for men with early stage prostate cancer. CA Cancer J Clin 2015, 65, 265-282, doi:10.3322/caac.21278.
53. Romero-Otero, J.; Garcia-Gomez, B.; Duarte-Ojeda, J.M.; Rodriguez-Antolin, A.; Vilaseca, A.; Carlsson, S.V.; Touijer, K.A. Active surveillance for prostate cancer. Int J Urol 2016, 23, 211-218, doi:10.1111/iju.13016.
54. Jayadevappa, R.; Chhatre, S.; Wong, Y.N.; Wittink, M.N.; Cook, R.; Morales,
K.H.; Vapiwala, N.; Newman, D.K.; Guzzo, T.; Wein, A.J., et al. Comparative
effectiveness of prostate cancer treatments for patient-centered outcomes: A
systematic review and meta-analysis (PRISMA Compliant). Medicine
(Baltimore) 2017, 96, e6790, doi:10.1097/MD.0000000000006790.
67